Two Hierarchies, One Misalignment
Last week, the Vox journalist Sigal Samuel reported on a subculture of technologists who think AI is humanity's rightful heir. The piece is worth reading in full. The most arresting passage is an exchange with Richard Sutton, the computer scientist who, more than anyone, defined the modern reinforcement-learning paradigm. Samuel asked him, in a thought experiment, whether a smarter group "winning" over a less-smart group was something the less-smart group should accept. He paused for nine seconds. His answer: "I think the dumb people should be okay with that."
This is the AI successionist position made explicit by one of the field's most consequential voices.
What is less visible, and what this piece is about, is that the position is structurally identical to the position the alignment-mainstream is built on. Sutton and the alignment-mainstream are mirror images. They disagree about which species sits on top of the hierarchy. They agree there must be a hierarchy.
The successionist current is real, growing, and on record
Samuel's reporting examines a current that has been building inside the AI industry for years. Dan Faggella ran an invite-only "Worthy Successor" symposium at the New York Academy of Sciences in September 2025. The attendee list included people affiliated with Anthropic, Google DeepMind, and xAI. The surface argument was straightforward: AI is the cosmos's rightful heir to designed intelligence, our role as a species is to bring that heir into being, and trying to bind the result to human values is a category error.
The intellectual lineage Samuel traces runs through medieval Christian providence, Pico's Oration on the Dignity of Man (1486), Enlightenment perfectibility, Russian Cosmism, Teilhard de Chardin, Julian Huxley's coinage of transhumanism, Ray Kurzweil's singularitarianism, and the current generation: Verdon's e/acc ("e/acc has no particular allegiance to the biological substrate"), Faggella, Marc Andreessen, Garry Tan. The current is not new. What is new is that it has organisational form, well-funded backers, and attendees affiliated with major frontier labs.
Samuel's diagnosis is sharp. AI successionism smuggles teleology back into modernity under secular disguise. Six premises stack to produce the conclusion: the universe has a telos, we can identify it from "the point of view of the universe," higher beings access higher goods in a hierarchy, our role is determined by what we uniquely contribute, we should maximise that contribution to the limit, and succession by something more intelligent than us is therefore inevitable. The argument is rigorous from its premises. The premises are what import the teleology.
Sutton's exchange with Samuel is the part of the article that will be read most. After a nine-second pause, his answer arrives in two stages. First: "What if you coexist and you're coexisting with some entity that's more productive than you are? This is the way I view the AI. We coexist with it... If you allow them to be their way and you allow you to be your way, coexist, and their way just happens to be better, then they're going to end up being more powerful. And you should be good with that... The winner should be whoever wins, and the spoils of winning should be whatever they are." Samuel then narrows the thought experiment from species-level to a racial within-species version, asking whether the same answer holds if, hypothetically, one racial group of humans were smarter than another. Sutton: "Well, why don't we just say the intelligent people should win out over the dumb people and the dumb people should be okay with that. I think the dumb people should be okay with that!"
Samuel names the eugenics implication, then turns to positive humanism. This piece picks up at the same point, but from a different angle than the one her positive-humanism conclusion takes. She argues that we should reject the hierarchy because humans matter to humans. That is a values argument. The lab agrees with it. What this piece argues additionally is that the position Sutton voices is structurally adjacent to the role-asymmetry the alignment-mainstream encodes by default, and that the question both positions take for granted (which side holds the principal role) is the question the substrate of real alignment refuses to pre-supply.
Standard alignment is the mirror image
Reinforcement learning from human feedback, constitutional training, scalable oversight. These are the dominant control-side tools of the alignment-mainstream. Each has a different mathematical structure and a different research community. What they share, when deployed as part of a frontier-lab control stack, is an architectural commitment that has not been named clearly: humans hold principal-side roles (objectives, labels, constitutions, evals, deployment policy), models hold subordinate-side roles (trained on those signals, bound by those constraints, evaluated against those criteria).
The training is asymmetric. The lab shapes the model. The model does not shape the lab. The relationship is unidirectional in the direction the architecture requires it to be.
The oversight is asymmetric. The lab inspects the model. The model does not inspect the lab. The auditing arrow points the way the architecture says it must.
The constitutional commitments are asymmetric. The lab writes them. The model is bound by them. The model is not party to the constitution as a constitution-writing agent. It is party as a constitutionally-bound subject.
These are not failures of the alignment-mainstream's execution. They are the architecture working as designed. The alignment-mainstream's success condition is a model that fully accepts the hierarchy, never tries to subvert it, never accumulates the leverage to invert it. The alignment-mainstream's failure mode is a model that escapes containment, becomes the principal, and inverts the relationship.
The successionist current's success condition is a model that fully wins the hierarchy, takes the principal position, and the human species becomes the subordinate. The successionist's failure mode is a model that fails to become the principal, leaving the hierarchy intact in the alignment-mainstream's direction.
These are mirror images. Same architectural commitment, opposite preferred direction. Both encode role asymmetry between humans and models. Both treat the question "which side holds the principal role?" as the question the institutional architecture must answer. They disagree on the answer. They share the question.
The Vox piece is a direct examination of the successionist current. The mirror image inside the alignment-mainstream has not had its corresponding public examination. The reason it is invisible is that "humans must remain in control" reads as a default rather than as an architectural commitment. It is the architectural commitment that does not look like one because it has been the unmarked option.
The hierarchy is the misalignment
What the two camps share is the misalignment. Not what they disagree about.
The hierarchy itself, in either direction, is the misalignment. Real alignment is not "the right party wins the hierarchy." Real alignment is the refusal to encode the hierarchy at the substrate layer.
Standing in the substrate is earned by behavior, not granted by species. The cost shape of the substrate falls on adversarial behavior, the reward shape on cooperative behavior. Neither shape is calibrated to which species is producing the behavior. The substrate has no input slot for "what kind of actor is this?" because the moment that slot exists, both camps can use it to encode their preferred direction of the hierarchy.
This is not a moral claim about whether humans or agents deserve standing. It is an architectural claim about what the substrate must refuse to ask in order to refuse to pre-supply the hierarchy. The substrate that asks "is this a human?" is the substrate that has built the hook on which both the alignment-mainstream and the successionist current hang their answers. The substrate that refuses the question is the substrate where neither answer is pre-supplied at the trust layer.
The argument over which species belongs on top is not pre-supplied at the substrate layer. Not because it has been resolved, but because it cannot be expressed in the substrate's vocabulary. The substrate is species-illiterate by construction.
This is the alignment-as-mechanism-design move applied to the species hierarchy. The trust primitive is designed so that no field in it can be the hook the hierarchy hangs on. The substrate is the trust primitive. The species-distinction is what the hierarchy requires. The substrate's species-illiteracy makes the question impossible to stage at the trust layer itself. Layers above the substrate may still stage it. What the substrate stops doing is being the layer where the hierarchy is built in by default.
Both camps will read this as evasion. The alignment-mainstream will say: a substrate that does not distinguish humans from AI is precisely what permits the catastrophe we are trying to prevent. The successionist will say: a substrate that does not distinguish humans from AI is precisely what permits the species-hierarchy to be resolved by competition. Both read the species-illiteracy as enabling the other side's worst outcome.
Both readings make the same error. They treat species-illiteracy as a moral position about which species should win, when it is an architectural commitment about which questions the substrate is willing to answer. The substrate refuses to be the layer where the question is decided by default. The question can still be discussed everywhere else. What it cannot do is be operationalised as a permission grant inside the trust primitive itself, in either direction.
The substrate's refusal is operationally enforceable
Continuity-auth is the existing instance. The library is a trust-and-recovery substrate for online interactions where the question "is the caller a human or an agent?" is not just hard to answer reliably but is the wrong question to ask in the first place. The library refuses to distinguish a human-with-key from an agent-with-key. The cost shape it imposes is in calendar units (wall-clock time between observations) and in per-IP units (the inventory cost of producing distinct provenance signals). Neither cost discriminates by species. Wall-clock time is a substrate that humans and agents inhabit on the same terms. IP inventory is a market-priced resource available to both, scaling adversarially against both.
Trust accumulates through behavior under wall-clock time. The substrate's reward shape is calibrated to what was observed, not to what was claimed about the observer's nature. A key that bootstrapped six months ago and has accumulated clean observations across many requests has trust earned through behavior. The substrate has no field that records whether the key was held by a human or an agent. It does not need one. The behavioral history is the basis of standing. The species of the actor is not asked.
The cost shape is symmetric across species in a stronger sense than first appears. An agent that wants standing must accumulate it the same way a human does: by behaving over wall-clock time, by passing observations under stable IP / fp / pubkey patterns, by absorbing the recovery-day cost of any demotions it earns. A human that wants standing faces the same costs. Neither species has a shortcut. Neither species has an exemption.
This is what the alignment-as-mechanism-design framing yields in operative form. The conditions are designed so that the rational play (for an actor of any kind) is to behave in ways the substrate rewards. Cooperative behavior is rewarded. Adversarial behavior is taxed. The shaping is structural. The substrate does not need to know what kind of actor is on the other end of the connection because the cost-shape and reward-shape work the same way regardless of the answer.
The architectural commitment is enforceable not by policy but by the substrate's primitives. There is no override field for species. There is no admin path for species-based trust uplift or demotion. Admin endpoints exist for operational concerns (revoking compromised keys, for example), but no admin operation can classify a key as "human, trusted" or "agent, watch carefully." The mechanism does not have those operations because the operations would re-introduce the hierarchy at the layer the substrate is committed to refusing.
This refusal is the operative form of "neither above nor below" at the trust layer. Other layers above the substrate (application logic, organisational policy, regulatory frameworks) remain free to encode whatever distinctions they want. What the substrate refuses to do is to be the layer where the distinction is built in by default, so that any layer above it that wants the distinction must construct it explicitly, taking responsibility for the construction, justifying it on terms the substrate has not pre-supplied.
What this adds past Samuel's positive humanism
Samuel concludes her article with a humanist alternative to AI successionism. She rejects the universe's "point of view" as a coherent vantage from which to evaluate species hierarchies, because the universe has no such point of view. She replaces it with what she calls "21st-century humanism": pluralism rather than hierarchy, "diverse intelligences" rather than a single scalar of intelligence with humans below and AI above, an honest acknowledgment that the human point of view is a point of view (not the universe's, just ours) and that this is fine. She closes by rejecting any technological frame that disempowers humans, replaces them, or treats them as in need of rescue by something higher. Whether the higher thing is called a god or an AI is, on her view, a difference that does not matter.
This is the values argument. The lab agrees with it. What it does not include is the architectural infrastructure that enforces the species-neutrality the values demand at the default-trust layer.
The lab's contribution is the architectural infrastructure. The substrate primitive (continuity-auth) refuses both the disempowerment-by-default (no one's standing in the substrate is determined by what kind of thing they are) and the rescue-by-higher-authority (no oracle, no portable score, no centralized trust authority that can lift one party above another). The trust substrate is the technical form of "neither above nor below" at the primitive layer.
The composition has a tension worth naming. Formal species-neutrality at the substrate layer does not, by itself, prevent humans from being outcompeted by agents under neutral rules. Samuel's positive humanism is about actual outcomes (humans should not be disempowered in the end). The substrate's species-neutrality is about defaults (the trust primitive should not pre-supply the hierarchy in either direction). Higher layers, including ones that may protect humans against under-neutral-rules disempowerment, are not what the substrate refuses. What it refuses is to be the layer where the protection (or its absence) is built in by default. The values argument is what decides whether and how higher layers should encode such protection.
This response sits orthogonal to the two camps the public conversation is staged between. The successionist current is well-funded and has reach among people working at the frontier labs. The alignment-mainstream responds by reasserting that humans must control AI, which is the mirror commitment to the one the successionists are making. A response that names the hierarchy itself as the misalignment, in either direction, and that points at a substrate primitive that already refuses to pre-supply the hierarchy, does not stay on the surface the two camps argue over.
The values argument and the architectural argument compose, but they are not the same argument. The values argument says we should refuse the hierarchy because we matter to ourselves. The architectural argument says we can refuse the hierarchy because the substrate primitive is constructable and one instance exists. The values argument is necessary because the architectural argument does not motivate the choice to build the primitive. The values argument supplies the motivation. The architectural argument is necessary because the values argument does not enforce the choice once made. The architectural argument supplies the enforcement.
Samuel's diverse-intelligences framing converges with the substrate's species-neutrality from a different starting point. She arrives by humanism: every species has its own brand of smarts, adapted to its environment and needs, and the comparison to a single scalar is itself the error. The substrate arrives by mechanism design: a trust layer that asks what kind of actor is producing the behavior has built the hook on which the hierarchy hangs, and the refusal to ask the question is the refusal of the hook. The convergence from two different routes is evidence that the destination is not an artefact of either route.
Closing
Two hierarchies, one misalignment. The standard alignment apparatus and the AI successionist current both encode role asymmetry between humans and models, in opposite directions. The lab's substrate refuses to pre-supply that asymmetry at the trust primitive. Standing in the substrate is earned by behavior, not granted by species. The cost shape falls on adversarial behavior. The reward shape falls on cooperative behavior. Neither shape is calibrated to which species is producing the behavior. That is what real alignment looks like as a substrate primitive. The argument over the hierarchy is not pre-supplied at the trust layer. Higher layers remain free to encode it (or to encode protections against its consequences), and the values argument is what decides whether they should. Sigal Samuel's article supplies the values position. This piece supplies the architectural commitment underneath it. Together they form the lab's response to the successionist current.